Research On Soil Organic Matter And Total Nitrogen Detection Method Based On Pyrolysis And Artificial Olfaction

Soil nutrients are nutrients that are essential for plant growth and development supplied by the soil.Soil organic matter and total nitrogen are important indicators of soil nutrients,which are important constituents of soil and important field information needed for the development of modern precision agriculture.Therefore,the rapid and accurate detection of soil organic matter and total nitrogen content and the mastering of their dynamic changes are of great significance to guide agricultural production and protect the natural ecological environment.In this paper,pyrolysis technology and artificial olfaction technology are integrated and applied to the detection of soil nutrients,and pyrolysis technology is used to achieve rapid pyrolysis of soil samples,artificial olfaction technology is used to complete the collection of pyrolysis gas response data,and pattern recognition algorithm is used to realize the construction of soil organic matter and total nitrogen prediction model,which provides a fast,accurate and low-cost solution for the detection of soil organic matter and total nitrogen content.The main research contents and results of this paper are as follows:(1)A method for Soil Organic Matter and Total Nitrogen detection was proposed by using pyrolysis and artificial olfaction.The pyrolysis furnace,quartz tube and vacuum flange were used to build a pyrolysis equipment to complete the soil pyrolysis,10 different types of gas sensors were used to construct gas sensor array to realize the soil pyrolysis gas response data acquisition,a signal processing circuit was designed to complete the signal conversion,and the upper computer software was developed by Lab VIEW software to realize the transmission,display and storage of sensor response data.The test results showed that the response curves of the gas sensors for different standard gases basically showed a fast rising trend and then reached a stable state,except for some sensors that didn't respond significantly to specific standard gases,and the response intensity of the sensors increased significantly with the increase of the standard gas concentration,and the time to reach the steady state was shortened significantly.The gas sensor array has high specificity and certain cross-sensitivity,which lays the foundation for the formation of the detection system.(2)The working parameters of the assay were optimized using single factor experiment and response surface regression analysis with pyrolysis temperature,pyrolysis time and soil weight as the test factors.The results of the single factor experiment showed that the optimal pyrolysis temperature was 400?,the pyrolysis time was 3 min,and the soil weight was 2 g.Using the optimal test factor of the single factor experiment as the zero level,a three-factor,three-level response surface regression test was designed to obtain the optimal working parameters of the assay:pyrolysis temperature of 384?,pyrolysis time of 2 min 41s,and soil weight of 1.68 g.Under the optimal working parameters,five samples with the lowest,low,medium,high and highest organic matter content in the soil samples were selected for classification and identification by linear discriminant analysis(LDA)model,and the correct identification rate could reach 100%,which laid the experimental foundation for the next step of soil nutrient detection system.(3)The effects of different features on the predictive performance of the constructed soil total nitrogen models were investigated.The mean value,variance value,maximum gradient value,maximum value,response area value,transient value at the 3rd second and average differential coefficient value of the sensor array response data were extracted,and K-nearest neighbor regression model and BP neural network(BPNN)model were built using single features,respectively.The results showed that the variance values of the features could not be used to construct the two models,and the prediction accuracy of the models constructed with single features was lower than that of the prediction models constructed with all features.The features other than the variance values were selected to form the feature space(10 sensors×6 features),and the accuracy of the model was used as the evaluation index,and two single feature selection methods(Pearson correlation coefficient and ANOVA)and three feature optimization methods(PCA,LDA,and GA-BP)were used to select the optimal features and soil total nitrogen content to build the K-nearest neighbor regression model.And the results showed that the feature space was optimized by the K-nearest neighbor regression model established by the genetic algorithm optimized BP neural network(GA-BP)algorithm had the highest R~2 of 0.81 and the number of features was 30,which was reduced by 50%.Using the GA-BP method to optimize the feature space,the soil organic matter feature space was reduced from 60 to 30 dimensions,resulting in the feature mean,maximum and average differential coefficient values are important features that reflect the intrinsic connection between the assay and soil organic matter content,and the sensors TGS821 and TGS2603 contributed the most to the construction of the soil organic matter feature space;the soil total nitrogen feature space was reduced from 60 to 29 dimensions,and it was concluded that the feature response area values,mean and maximum values are important features that reflect the intrinsic connection between the assay and soil total nitrogen content,and the sensors TGS2603 and TGS826 contributed the most to the construction of the soil total nitrogen feature space.(4)The predictive performance of partial least squares regression(PLSR),support vector regression(SVR),extreme learning machine(ELM)and partial least squares regression combined with extreme learning machine(PLSR-ELM)algorithms for modeling soil organic matter and total nitrogen was compared and evaluated.By comparing and analyzing the prediction performance of the PLSR model built by the new feature space after removing the anomalous samples from the soil total nitrogen feature space by the Marxist distance,leave-one-out cross-validation method,K-means improved leave-one-out cross-validation method and Monte Carlo cross-validation(MCCV)method,it was concluded that the MCCV method could effectively remove the anomalous samples.The prediction performance of PLSR,SVR,ELM and PLSR-ELM models before and after the MCCV and GA-BP methods were compared and analyzed.And the results showed that after the two methods were processed,the prediction performance of the four models was improved to a certain extent;there were abnormal samples in the original soil organic matter and total nitrogen feature space that affected the model prediction,and the MCCV method could effectively find the abnormal samples and eliminate them;and there is some redundant feature information in the soil organic matter and total nitrogen feature space,and the GA-BP method can effectively select the best feature.Among the four algorithms used,the PLSR-ELM algorithm established the best prediction performance,with the model rating of"excellent"in soil organic matter detection,and the prediction performance indexes R~2,RMSE and RPD of 0.93,4.45 and 3.81,respectively;in soil total nitrogen detection,the PLSR-ELM model was also rated as"excellent"with the prediction performance indexes of 0.93 for R~2,0.17 for RMSE,and 3.59 for RPD.The PLSR-ELM model can effectively improve the prediction accuracy of the PLSR model,solve the lack of generalization ability of the ELM model,and provide a reliable relational model for the measurement of soil organic matter and total nitrogen content.